A GitHub repository, usually shortened to a repo, is a project folder hosted on GitHub that holds the code, its full history, the README, the issues, the releases and the licence.

Anyone can view and download a public repository. Cloning one gives you the entire project along with every change that has ever been made to it.

The best GitHub repos for AI fall into a handful of jobs. Agent frameworks give you the loop, MCP servers give an agent real capabilities, and the rest covers coding agents, prompt libraries, RAG and memory, and eval harnesses.

The list on this page is filtered by exactly those categories, and every repository on it is one we have installed or read properly rather than something pulled off GitHub trending.

Sort by what is maintained rather than by stars. Check the last commit date, open the issues tab to see whether anyone answers, and read the first paragraph of the README, because a README that does not tell you what the project does is telling you something.

GitHub search, topics pages and trending are all reasonable starting points. Curated lists like this one exist because trending rewards novelty rather than usefulness.

There is no difference. Repo is simply the shorthand, and it is what almost everyone says out loud.

GitHub uses the full word repository across its own interface, so you will see both spellings in the docs and on the site.

No, and this catches people out constantly, because public and open source are not the same thing.

A repository is open source only if it carries an OSI-approved licence such as MIT, Apache-2.0, GPL or BSD. Some popular projects are source-available instead, under licences like BSL, Elastic, SSPL or Commons Clause, which means you can read the code but you cannot build a competing commercial product on it. A repository with no LICENSE file at all is legally all rights reserved, so you may read it and do nothing else with it.

Use the licence filter above to see which is which.

Everything listed here is free to download, and the open source projects are free to use in your own work within the terms of their licence.

Some of them are the open core of a paid hosted product, which means the code is free while the managed version is not. Where that applies, the drawer links out to the project's own site.

MCP stands for Model Context Protocol, and an MCP server gives an AI agent a real capability it does not have on its own, such as reading a database, driving a browser, calling an API or editing files.

Install one and your agent can do the thing rather than describe it. We keep a full library of them at AI connectors.

Stars measure attention rather than quality, and they never go down, which is why a repository that was abandoned three years ago can still show forty thousand of them.

Treat the number as a rough popularity signal and weight the last commit date, the open issue count and the number of contributors far more heavily. A project with nine hundred stars that ships every week is usually the better bet.

Read the README first and follow its install instructions, because in many cases a single command is all it takes. Where one exists, we put that command in the drawer as a copy block.

To get the code itself, run git clone followed by the repository URL, or download the ZIP from the green Code button.

Star counts, forks, language, licence and last commit date come straight from the GitHub API and refresh every day, so the numbers here are never more than twenty-four hours old.

The write-ups are ours and we revise them whenever something material changes about a project. We add new repositories most weeks.

Someone here has to have used the project or read it properly, and that is the entire bar, which is why this list is short rather than exhaustive.

If a repository you rely on is missing, tell us about it and say what you built with it.